Time-to-Event Predictive Modeling for Chronic Conditions Using Electronic Health Records

Abstract

Although electronic health records (EHRs) hold promise for supporting clinical decision making, few studies have used them to model the progression of chronic conditions. To examine the feasibility of EHR-based predictive models for chronic conditions and to mitigate the associated data challenges, the authors develop a time-to-event predictive modeling framework consisting of five analytical steps: guideline-based feature selection, temporal regularization, data abstraction, multiple imputation, and extended Cox models. Using concept- and temporal-abstracted features, the proposed model attained significantly improved performance over the model using only base features.